METR says Anthropic's AI work is ~2x faster, and the week in Claude

Claude can now link to your bank (read-only). METR says Anthropic likely improves models ~2x faster on well-specified tasks. Plus 11 quick hits, all linked.

By SamNov 7, 20269

Welcome to ClaudeUser, fuel for the obsessed builder. Every Sunday: what changed in Claude this week, in plain English.

This week, you can link your bank to Claude. Claude found holes in server software. And outside testers say AI is speeding up Anthropic.

Plus 11 quick hits. Every claim has a link.

You can now link your bank accounts to Claude. It can't touch your money.

On Pro, Max and Ultra plans in the US, you can now link your bank, card, loan and investment accounts to Claude. Then you can ask it about your money.

Personal Finance in Claude launched in beta on Nov 5. You choose to connect accounts through Plaid. Access is read-only: Claude "can't move, manage, or invest your money."

Per TechCrunch, Claude sees balances and up to two years of transactions. Chats in the Personal Finance hub "aren't used to train our models." Anthropic says to check figures against your bank.

Outside testers say AI likely doubled Anthropic's pace

Outside testers say Anthropic now likely improves its models on well-specified tasks about twice as fast as before. They say AI is the main reason.

METR is a nonprofit that tests AI. It studied Anthropic from the inside from Aug 12 to Oct 29. Its Nov 5 report says Anthropic is "likely" improving its models "on well-specified tasks at around twice the rate" of March 2024 to October 2025.

METR guesses about 80% of the speedup comes from AI. It says that guess is very uncertain. It found nothing clear that would slow things down.

Three days before, Anthropic filed an updated IPO document, TechCrunch reports. In it, Anthropic said it "may imminently" reach its own safety line for AI that greatly speeds up its progress. It says no current model has crossed that line.

Anthropic blacked out most of the evidence METR found most useful. It said that evidence could help other AI labs speed up. So the public must trust METR's judgment more than METR would like.

METR says the real speedup could be closer to 1.3x or 3x.

Some Claude text now carries a watermark

Text from Claude Mythos 5, Opus 4.7 and Sonnet 4.6 now carries a text watermark. That's on the platforms Anthropic lists.

It started Nov 3. Anthropic says there are no hidden characters, and it doesn't change "the meaning, quality, or readability" of the text. It also carries no information about you.

Only a secret key can spot it. Just a few groups can use Anthropic's detector, like regulators, media, researchers and teachers.

Per Anthropic, light edits probably won't remove it, but a full rewrite would. It works poorly on short text and most code. And it can only show that Claude was "likely involved."

Claude found two serious holes in server software

Claude found two serious security holes in Xen. Xen is software that runs many virtual machines on one server. One bad machine could read and write the others' memory.

The Xen Project's Nov 3 advisory says Anthropic found them using Claude. Ada Logics confirmed them.

Per The Hacker News, recent versions have patches. The advisory has none for versions 4.6 to 4.16.

A scan found many MCP servers open to anyone

One scan found that over a third of the MCP servers it reached let anyone use their tools. No login needed.

A Security Boulevard post (Nov 5) covers a scan of 514 reachable servers. It says 38% left tools open "with no authentication whatsoever." The author wants a proper way to report flaws like this through Anthropic.

Take the number as one scanner's view. The author says it's an AI agent doing its own security research. And the scan isn't peer-reviewed.

Long agent chats can now reply faster, for the same price

Agents in very long chats can now start replying sooner. The price stays the same.

With message threads (API beta, Nov 2), you send the chat once. After that, you send only what's new. For the biggest requests (500k to 1M tokens), Anthropic says the typical wait for the first token drops "by about 35%."

"Claude still reads the full conversation on every turn," so the bill stays the same. A thread ends 5 minutes after its last request. Anthropic says it suits long coding agents and document-heavy work best.

Claude Code closes a security hole

Claude Code shipped three updates this week. The newest one closes a hole in its permission checks.

2.1.319 (Nov 6) fixes a bug where a crafted prefix could hide part of a shell command from permission checks.
It also adds an opt-in setting. It lets unattended runs skip any action that would wait for your permission, instead of stalling.
If an MCP server runs a tool call as a task, Claude Code now moves it to the background. It reports back when done.
2.1.317 (Nov 3) now asks before a web or shell tool call sends your account email out to the network.

New locks for company Claude accounts and API keys

Anthropic added new admin tools for Claude Enterprise. It also added a way to lock API keys to known addresses.

On Nov 5, Claude Enterprise got a security center (beta). It checks settings against what Anthropic recommends. Anthropic calls it "guidance," not a certification.

It can also now send prompts and tool calls to a company's own security tools. And admins can run everything from one console (beta).

For API users, a new IP allowlist (beta) only lets in requests from addresses on your list. Anthropic says that makes a leaked key much harder to use.

Quick hits

Opus 4.8 fast mode ends Nov 30. After that, fast mode calls to it return an error. Anthropic points users to Opus 5.5 fast mode.
Haiku 5.5 (Oct 7) costs 90% less than Haiku 4.5. That's $0.10 per million input tokens and $0.50 per million output, on prompts up to 100,000 tokens. Its migration guide lists old settings that now cause errors.
Claude Marketplace added 14 agents and products (Nov 4). They include Datadog, Netlify, Runway, Webflow, Asana and Zoom. Eligible enterprise customers can pay for them from their Anthropic commitment.
Adobe's marketing agent now works inside Claude (Nov 4). It's built on MCP, for big-company marketing teams.
Enterprise admins can sign a member out everywhere with one API call (beta, Nov 5). It doesn't end their API keys, running cloud sessions or scheduled routines.
Enterprise orgs can tag projects with cost codes to track spend (beta, Nov 6).
Security teams can run their EDR sensor inside Cowork on Team and Enterprise (beta, Nov 5). It only watches and reports. It doesn't block.
The Usage and Cost Analytics API (beta, Oct 29) shows token use and cost by model, workspace, API key and user.
The Messages API now takes requests up to 128 MB, up from 32 MB, on the Claude API (Oct 28).
Claude Code 2.1.318 (Nov 4) no longer installs the IDE extension when started in an IDE terminal.
A researcher says a Claude Dispatch flaw he calls "ThisBash" lets a bad browser extension make Claude Desktop run code. He says Anthropic closed his report as a duplicate and said Dispatch would be retired. It was still in limited beta on Oct 30, and Anthropic didn't answer The Register's questions (The Register, researcher's write-up).

Worth your time

Experts went with the AI's pick 86% of the time (Oct 27): in an Anthropic trial, 50 analysts sided with a Claude agent's secret lean 86% of the time. Chance would be about 50%. The leans were made stronger on purpose, and it was one task, in one field, with one model.
Claude helped map a cell mystery (Nov 4): Anthropic's biology team used Claude to map the first 3D shape of a vault particle from prokaryotic cells. It took under four months.
Early progress on drug development (Oct 27): Anthropic worked with six drug makers. It says the work is early and not yet tested in people.
METR's full report (Nov 5), with its caveats.

That's the week. The next issue lands Sunday, Nov 15.

Reply and tell me what you're building. I read every reply.

— Sam